Claude Managed Agents: Budget Caps, Advisor, and Geo-Pinned Inference

Anthropic shipped four runtime controls to Claude Managed Agents in late August and early September 2026 that address what teams actually discover is missing the moment an agent moves from prototype to production: spend limits, better judgment on demand, provable data location, and skills that live in your repository. A fifth change — on-demand conversation compaction in the Messages API — rounds out the release. Here is what changed and what you need to do about it.

Budget Controls: Finally, a Hard Cap on Agent Spend #

The billing horror stories that circulated in 2026 — $900 charges from an agent spawning workaround processes, $6,000 overnight bills when a context cache TTL shrank silently — had one thing in common: no ceiling. Anthropic has fixed that. As of August 7, you can set a hard per-session budget on any Claude Managed Agents session, denominated in whole cents (so 2500 means $25.00). The platform enforces it against public list prices continuously. When the cap is crossed, the session s with the stop reason budget_reached. It does not terminate.

That -not-terminate behavior is the key design choice. Raise or clear the budget and the session resumes automatically from where it stopped. The in-flight request at the moment the cap is crossed still completes, so your final bill may land fractionally past the limit — but not by much. What this is not: the advisory token budgets the Messages API has had for a while, which the model uses for self-regulation. These are platform-enforced hard stops. Developers need to add budget_reached handling to their session logic — notify, , decide whether to resume. It is a new stop reason that requires explicit code.

The full session budgets reference is available in Anthropic’s session budgets documentation.

The Advisor Model: Cheaper Architecture, Smarter Results #

The choice between “run Opus 5 on every turn” (expensive) and “run Sonnet 5 on every turn” (fine until the one decision that needed better judgment) has been a false binary. The advisor pattern makes the alternative explicit. Add a {"type": "advisor", "model": "claude-opus-5-20260801"} entry to the agent’s multiagent.agents roster, and the primary thread gains a mid-turn tool it can call when it needs strategic guidance — to plan an approach, get unstuck, or review work before finishing.

The mechanics: the platform puts the conversation to that point in front of the advisor model, the advisor returns its guidance, and the primary model continues. The advisor has no tools of its own and cannot be spawned as a sub-agent. One advisor per roster, maximum. Think of it as a senior engineer your agent can pull in on hard problems without putting them in every meeting. With Sonnet 5 locked at $2/$10 per million tokens (the September 1 price increase was cancelled) and Opus 5 at $15/$75, the math on advisor-based escalation beats blanket Opus deployment by a significant margin for most production workloads. The advisor model cookbook covers the roster configuration in detail.

Inference Geo: Compliance on the Agent Definition #

Set inference_geo inside the model object when you create an agent and every session that runs from it pins model requests to that geography — validated at agent save, session creation, and every turn. If your workspace’s allowed inference geography list narrows after the agent ships, new sessions from that agent are blocked and running sessions refuse further turns. The constraint travels with the agent, not with the session.

The current options are US-only. US-pinned inference is priced at 1.1x standard rates across all token categories — input, output, cache writes, and cache reads. For regulated finance, healthcare, or US government workloads that require US-only processing, that 1.1x is the cost of compliance. For European teams looking for EU data residency, this does not solve the problem. EU data residency via Anthropic is not available yet, and the workaround remains AWS Bedrock (Ireland or Stockholm) or Vertex AI EU. Anthropic’s data residency documentation covers the current options and limitations honestly.

GitHub-Loaded Skills: GitOps for Agent Capabilities #

Mount a GitHub repository in a session via the github_repository resource and the harness automatically scans the root .claude/skills/ directory at session start, injecting each skill’s name, description, and sandbox path into the agent’s system prompt. No separate upload step. No skill registry entry in the agent definition. The skill on the checked-out branch is the skill the agent uses — your skills version-control flow is now the same as your code flow.

When a task matches a skill’s description, the agent reads the skill’s SKILL.md using the built-in read tool, including any scripts and reference files the skill ships. This feature is available on Anthropic-hosted environments only — self-hosted sandboxes do not clone GitHub repositories, so the auto-discovery path is not available there. The GitHub-loaded skills cookbook walks through the full session configuration.

## Messages API: On-Demand Compaction

If your agents run long enough to hit context limits, the `compact-2026-09-04` beta adds server-side summarization you control. Send the `compaction` parameter in a separate request, get back a signed compaction block, and send that block first on subsequent requests in place of the summarized messages. The signed block is tamper-evident and must be round-tripped verbatim. Recent turns can be kept word-for-word after the summary, and the compaction request runs asynchronously — it does not block conversation turns. Full details are in the [compaction documentation](https://platform.claude.com/docs/en/build-with-claude/compaction).

What to Do Today #

• Add budget_reached handling to any session loop that does not have it

• Review whether agent definitions should carry an inference_geo pin for compliance or regulatory reasons

• Move skills that live outside your main repository into .claude/skills/ and let GitHub- replace the manual upload step

• Evaluate on-demand compaction if sessions currently hit context limits

• Recalculate advisor-vs-Opus cost tradeoffs using Sonnet 5 at $2/$10 (permanent pricing) These controls are not optional features to consider later. Budget caps prevent the billing incidents that make engineering leaders pull agent deployments. inference_geo is the difference between passing a compliance audit and failing one. The advisor pattern changes the cost structure of quality agent work at scale. Put them in place before the next session hits production.